whisper-medium-ct2

OpenAI Whisper medium in CTranslate2 format, converted and published by Diction Labs for self-hosted speech-to-text.

Use with Diction

Self-hosted Diction serves this through Speaches. Point the gateway's custom backend at your Speaches container and name this repo as the forwarded model. Both variables are required: the custom backend only activates once the URL is set.

environment:
  CUSTOM_BACKEND_URL: http://whisper-medium:8000
  CUSTOM_BACKEND_MODEL: DictionLabs/whisper-medium-ct2

Full stack in the self-hosting guide.

Use directly

from faster_whisper import WhisperModel
model = WhisperModel("DictionLabs/whisper-medium-ct2")
segments, info = model.transcribe("audio.mp3")

Conversion

Converted from openai/whisper-medium with:

ct2-transformers-converter --model openai/whisper-medium \
    --output_dir whisper-medium-ct2 --copy_files tokenizer.json --quantization float16

Weights are FP16. Change it at load time with CTranslate2's compute_type.

Built with CTranslate2 4.8.1, which writes vocabulary.json. Builds from CTranslate2 3.x shipped vocabulary.txt instead. Any faster-whisper on CTranslate2 4.x reads this layout; verified against faster-whisper 1.2.1.

Published under our own namespace so self-hosted Diction stacks pull from a repository we control, converted from OpenAI's checkpoints rather than repackaged from another distribution.

MIT licensed, as is Whisper itself.